Add braided evolution prompt (BUILD-TEST alternation for deeper mutation plans)

[Cranot]

Summary

• Adds BRAIDED_EVOLVER_SYSTEM_PROMPT to adaptive_skill/prompts.py -- a 5-step BUILD/TEST alternating prompt for evolution analysis
• Adds _system_prompt property to AdaptiveSkillEngine that selects prompt based on config.extra["evolver_style"]
• Default behavior unchanged -- braided prompt only activates with evolver_style: "braided"

What it does

The default evolution prompt analyzes failures then generates fixes (BUILD-BUILD-BUILD-BUILD). The braided prompt alternates structure-building and structure-testing:

Step 1 [BUILD]: Analyze failures, name conservation law (structural trade-off)
Step 2 [TEST]:  Challenge analysis -- what does it conceal? Would fixes break passing tasks?
Step 3 [BUILD]: Cross-reference failure categories, find interaction patterns
Step 4 [TEST]:  Audit existing skills -- helping or hiding the problem?
Step 5 [BUILD]: Mutation plan with root-cause reasoning + regression predictions

Evidence

Tested on synthetic observation data (8 failed tasks across 4 categories, 3 existing skills, 60% pass rate):

Condition	Score	Root Cause	Self-Challenge	Cross-Ref	Skill Audit	Prediction	Actionability
Default prompt	6.4	3	2	1	2	2	3
Braided prompt	10.0	5	3	3	3	3	3
Delta	+3.6	+2	+1	+2	+1	+1	0

Scored by a custom mutation plan evaluator (6 dimensions, 0-20 scale normalized to 10). The default produced category-level triage with per-category patches. The braided prompt found a unified root cause (working-memory decay under long trajectories) that the category-level analysis concealed, and identified an existing skill that was masking the problem.

Usage

config = EvolveConfig(extra={"evolver_style": "braided"})
engine = AdaptiveSkillEngine(config)
# Everything else unchanged -- same workspace, observations, history, trial

Or in YAML config:

evolver_style: braided

Design principles

Based on research from the AGI-in-md project (330 principles on cognitive compression in LLM prompts, 1000+ experiments). Three principles applied:

• Braid balance (P303): Alternating build/test phases outperform monotonic sequences. Empirically: braid-balanced avg 8.2 vs monotonic 6.5 (+1.7) across 50+ experiments.
• Conservation law anchoring (P305): Naming structural trade-offs ("when X increases, Y must decrease") produces categorically deeper analysis than listing problems.
• Cross-analytical reference (P315): Analyzing one layer's findings through another's lens discovers interaction patterns invisible to parallel independent analysis.

Test plan

• Import verification (BRAIDED_EVOLVER_SYSTEM_PROMPT loads correctly)
• Default behavior unchanged (no evolver_style = uses DEFAULT_EVOLVER_SYSTEM_PROMPT)
• Braided selection works (evolver_style: "braided" = uses BRAIDED_EVOLVER_SYSTEM_PROMPT)
• End-to-end benchmark comparison (would welcome help testing on actual SWE-bench/MCP-Atlas)

Scope

Minimal change: 2 files, +62/-3 lines. No new dependencies. No breaking changes. The braided prompt is opt-in only.

[HanqingLu]

Since the adaptive_skill has been used by other people to run other benchmarks, in order for this to be a backward compatible changes, suggest to clone a new algorithm and reimplement. Also, could you please share the benchmark numbers (not on your test but on the assigned benchmarks like Terminal-Bench-2, SWE-Bench?